towards an artiﬁcial olfactory mucosa for improved odour classiﬁcation · 2008-10-03 ·...

Towards an artificial olfactory mucosafor improved odour classification

BY JULIAN W. GARDNER1,*, JAMES A. COVINGTON

1,

SU-LIM TAN1

AND TIMOTHY C. PEARCE2

1School of Engineering, University of Warwick, Coventry CV4 7AL, UK2Department of Engineering, University of Leicester, Leicester LE1 7RH, UK

Here, we report on a biologically inspired analytical system that represents a newconcept in the field of machine olfaction. Specifically, this paper describes the designand fabrication of a novel sensor system, based upon the principle of ‘nasalchromatography’, which emulates the human olfactory mucosa. Our approach exploitsthe physical positioning of a series of broadly tuned sensors (equivalent to the olfactoryepithelium) along the length of a planar chromatographic channel (analogous to thethin mucus coating of the nasal cavity) from which we extract both spatial (responsemagnitude) and temporal (retentive delay) sensor signals. Our study demonstrates thatthis artificial mucosa is capable of generating both spatial and temporal signals which,when combined, create a novel spatio-temporal representation of an odour. We believethat such a system not only offers improved odour discrimination over a sensor array-based electronic nose, but also shorter analysis times than conventional gaschromatographic techniques.

Keywords: electronic nose; artificial olfaction; chemical sensors; olfactory mucosa

*A

RecAcc

1. Introduction

It is not uncommon for an increase in the understanding of a biological system tolead to a significant technological advance. In the evolution of the science ofmachine olfaction, and ‘electronic nose’ (e-nose) technology, there have beenmilestones in the areas of sensor design, sensing materials and packaging(Gardner & Bartlett 1999). These past developments have led to thecommercialization of e-nose instruments and their subsequent application in abroad range of industries, including food, environmental and healthcare (Pearceet al. 2003). Traditional e-noses operate by making use of a small number ofchemical sensors with partially overlapping sensitivities. Each sensor is tuned toa different chemical group so that when an array of sensors is presented with acomplex odour its chemical components produce a response vector or odourfingerprint. Although many instruments are commercially available today, theirability to identify complex odours in real world situations is still somewhatlimited. This shortcoming is further highlighted when such instruments are

Proc. R. Soc. A (2007) 463, 1713–1728

doi:10.1098/rspa.2007.1844

Published online 24 April 2007

uthor for correspondence ([email protected]).

eived 24 November 2006epted 13 March 2007 1713 This journal is q 2007 The Royal Society

J. W. Gardner et al.1714

compared to their biological counterpart, which is not only able to identifyodours at lower (by orders of magnitude) concentrations but also within abackground of other time-varying interfering odours.

Greater system complexity and redundancy in the biological system over andabove existing technological implementations are responsible, at least in part, forits superior operating performance. The human olfactory system, for instance,has some 100 million olfactory receptors (ORs) with approximately 350 differenttypes of receptor proteins expressed by the human genome in the olfactoryepithelium (Malnic et al. 1999; Goldstein 2001). In contrast, an e-nose has asmall number (typically less than 32) of chemical sensors in a basic chamber anda chemometric data processing algorithm. Until now researchers haveconcentrated predominantly on deploying different sensing materials and theirchemo-reception mechanisms to approach the impressive diversity andsensitivity of ORs; however, few have researched either the physiology of thebiological system to study whether it offers any advantage or studied carefullyhow odours should be optimally delivered to an olfactory sensing device. Here,we propose to mimic the basic structure of the input stage of the human olfactorysystem, specifically, emulating the properties of the olfactory mucosa.

The olfactory mucosa lies at the top of the human nasal cavity (within thesuperior turbinate) and comprises the olfactory epithelium and mucus layer—figure 1(a). This olfactory epithelium contains the receptor cells and hair-likecilia structures with the different olfactory binding proteins, which aredistributed along and beneath the thin mucus layer. Various studies havedemonstrated the mucus layer coating the nasal epithelium has partitioning (orgas chromatographic; GC) type properties that contributes to the coding ofolfactory information (Mozell et al. 1987; Kent et al. 2003). Receptor cellsdistributed beneath this ‘nasal chromatograph’ have been shown to generate bothspatial1 (response magnitude) and temporal (time-delayed) signals (Kent et al.1996), which are thought to give rise to the generation of a spatio-temporal mapof the odourant being generated and conveyed to the olfactory bulb. It has beensuggested that these signals could, plausibly, be an important factor in enhancingour ability to discriminate between similar odours (Quenet et al. 2001; Lysetskiyet al. 2002).

Such a segregated approach to odour delivery has yet to be exploitedtechnologically but potentially offers performance gains from the spatio-temporaldynamics of the sensor population response; thereby contributing informationabout the stimulus. We believe that this approach will lead to significantimprovement in the capability of a new generation of e-noses. For instance, such asystem would offer faster analysis times over conventional GC-based methodsand be able to identify simple and complex odours better than current e-noseinstruments. Moreover, segregation of odours at the front end of the olfactorypathway may be a key factor in solving complex odour discrimination tasks wherethe system needs to be sensitive to the relationship between individual odourcomponents existing within a blend. The challenge here is to determine how bestto exploit this approach, i.e. by combining it with modern fabrication technologiesto create a new architectural platform for future e-nose implementation. In this

1 For the sake of expediency, we decided to adopt the term ‘spatial’ here to describe the maximumsignal in odour space and to differentiate it from the simple time-delay (temporal) signal.

Proc. R. Soc. A (2007)

mid-sagittal view of head

impulse

olfactory bulb

cilia

olfactory (receptor) cell odourant

olfactory mucosa

end of channel

sensor

simplified equivalentstructure

temporal sensorseparation

spatial sensorseparation

retentive coating

odour flow

(a)

(b)

Figure 1. (a) Mid-sagittal view of a head showing the location of the olfactory mucosa (circled)with a detailed cross-section for part of the human olfactory mucosa showing the mucus layer andreceptor cells and (b) our simplified equivalent structure of the artificial mucosa.

1715Towards an artificial olfactory mucosa

paper, we report on our initial effort to develop an artificial mucosa system,employing finite-element modelling, device fabrication and odour testing, in orderto provide superior odour discrimination.

2. Artificial olfactory mucosa design and modelling

Our artificial mucosa comprises multiple chemical sensors distributed along thelength of a gas chromatograph channel, as shown in figure 1b. GC separationrelies on a layer of retentive material (stationary phase) coating the inner surfaceof a column or channel. When a fixed volume of compound analyte (e.g. complexodour) is injected into the column with a carrier gas, each component withinthe compound analyte travels along the column at a different velocity



(mobile phase), depending upon its affinity with the stationary phase. Hence,towards the end of the column, a sequence of pulses will elute representing theseparated chemical components, depending on its length.

A finite-element model (FEM) was developed to investigate the charac-teristics of the artificial olfactory mucosa. A commercial FEM package (FEMLAB,Comsol, UK) was used to simulate analyte transportation and retentionbehaviour. A multi-physics model, employing the Navier–Stoke equation, wasused to simulate laminar transport. Convection and diffusion models were usedto simulate the diffusion and dispersion of the injected analyte pulse (Fonteset al. 2001). Diffusion was used to model the dispersion and partitioning effect ofthe analyte (Tan et al. 2006) in the coating. Partial differential equations wereused for the simulations and have been described in detail elsewhere (Tan 2005).To ensure that the model was an accurate estimate of the actual phenomena,simulations of equivalent GC columns using a standard commercial GC softwarepackage were performed, hence cross-validating results with establishedanalytical solutions. The mucosa model (1.2 m long column with 500 mm2

cross-sectional area and 10 mm coating layer) was found to perform similarly toa GC column with errors of less than 3% in retention time and less than 9% inseparation factor.

The real olfactory mucosa is relatively short (approx. 10 cm) when comparedwith traditional GC column lengths (approx. 10 m). Moreover, the biologicalsystem has analysis times that are drastically quicker (much less than 10 s) thanboth GC systems (approx. 10 min) and traditional e-nose instruments (approx.1 min). Hence, we have chosen a reduced channel length to be closer to thehuman olfactory system rather than requiring complete component separation.Thus, our analysis will be performed on information that can be extracted fromthis shorter channel rather than seeking perfect GC-like component separation,as is likely to be the case in biology. Our artificial mucosa comprised a 2.4 m!0.5 mm!0.5 mm long rectangular channel with 40 discrete sensors of fivedifferent tunings (i.e. five different sensing materials with differing specificities).A 10 mm thick layer of Parylene C (polymonochloroparaxylylene C) was used asthe retentive coating. The use of this material as a stationary phase was firstintroduced by Noh et al. (2002) and it has been shown that this material behavesas a stationary phase coating. The 40 sensors were placed into eight groups of fivewith each group containing the five different tunings. Details of sensor typesand their placement along the channel are given in table 1. Simple polar andnon-polar odours, namely ethanol and toluene, were used in the computersimulations in order to compare and optimize the operation of coated anduncoated channels.

Simulation results showed that the optimum flow velocity for columnefficiency was approximately 7 cm sK1 for a 5 s pulse of ethanol and toluenevapour in air, although the effect of increasing flow velocity was found to bemarginal on column efficiency. This velocity gave the best compromise betweenreducing the unwanted pulse broadening in which the faster the flow is the lessthe odour pulse front has time to diffuse out while traversing down the channel,and increasing the desired retention effect. In the latter mechanism, the lower theflow velocity is, the better is the separation of chemical components for acomplex odour. These simulations give the analyte concentration profiles at thedifferent physical locations rather than the corresponding sensor responses.


Table 1. Sensor placements within the artificial olfactory mucosa for both the simulations andfabricated system. PEVA, polyethylene-co-vinyl acetate; PSB, polystyrene-co-butadiene; PEG,polyethyleneglycol; PCL, polycaprolactone; PVPH, poly 4-vinyl phenol.

no. distance (mm) type no. distance (mm) type

S1 10 PEVA S21 1260 PEVAS2 30 PSB S22 1280 PSBS3 50 PEG S23 1300 PEGS4 70 PCL S24 1320 PCLS5 90 PVPH S25 1340 PVPH





In order to obtain the complete system model, the sensor responses werecoupled to the analyte profiles. For simplicity, an analytical method (defined byequations (2.1) and (2.2)) was used to model the dynamic response characteristicsof previously tested chemical sensors. Both the finite-element simulations and thefabricated system used polymer composite sensing materials, in a resistiveconfiguration, as the odour-sensitive layer. These composites comprise electricallyconducting carbon nanospheres within a non-conducting polymer binder. Thus,the carbon endows measurable electrical conductivity to the resultant compositematerial. These sensing materials are believed to operate through a chemo-mechanical effect, where the reversible absorption of an analyte results in aphysical swelling of the polymer. The swelling reduces the mobility of electronshopping between nanospheres within the bulk material and thus increases theelectrical resistance of the film. These sensing materials were chosen owing to theirrapid (approx. 100 ms) response time, ease of deposition, room temperatureoperation and the wide variety of available polymers. The response magnitude andresponse time, to pulses of ethanol and toluene vapour in air, were measured bythe five polymer/carbon black composite sensors (mixed 80 : 20 by weightpolystyrene-co-butadiene (PSB), polyethylene-co-vinyl acetate (PEVA), poly-ethyleneglycol (PEG), polycaprolactone (PCL) and poly 4-vinyl phenol (PVPH)).Experiments were carried out at different flow velocities (up to 1600 cm sK1), at atemperature of 30G28C, and relative humidity of 40G5%, while the responses(defined as the change in resistance from the baseline value in air) were measured.


Table 2. Values of model ‘on’ parameters estimated by fitting experimental data from PSB, PEVA,PEG, PCL and PVPH composite sensor responses to pulses of ethanol and toluene vapour in air.R is the Pearson correlation coefficient with a perfect fit giving a value of one. Fitted values forthe ‘off’ parameters are not employed in the data analysis performed here but may be found inTan (2005).

sensor type

ethanol toluene

RON tON R2 RON tON R2

PSB 0.5346 0.3568 0.8638 3.3282 0.4062 0.9902PEVA 2.6793 0.7229 0.9626 20.9653 0.5349 0.9870PEG 35.1677 0.0630 0.9945 32.5009 0.1445 0.9948PCL 0.8580 0.103 0.9477 7.7427 0.0807 0.9964PVPH 2.1748 0.1887 0.9928 2.1630 0.1881 0.9950


These were carried out in a purpose-designed micro-chamber system to help createsuitable boundary conditions for simple laminar plug flow. Details of thecharacterization and test systems are described in Tan (2005). A number ofstudies have suggested that the response of these sensors can be approximated to abilinear exponential model (Gardner et al. 1995, 1999). Using this simple model,the response of a sensor to a step increase in odour concentration c or ‘on transient’is given by

Rðc; tÞZRONðcÞ½1KeKtONt�; ð2:1Þwhere RON is the maximum response magnitude (i.e. at odour concentration co)and tON is the characteristic rise time. Conversely, the ‘off transient’ is given by

Rðc; tÞZROFFðcÞeKtOFFt; ð2:2Þ

where ROFF is the decay magnitude and tOFF is the characteristic decay time. Inpractice, RON and ROFF have very similar magnitudes, whereas tOFF is usuallymuch larger than tON. Table 2 gives the values of the different ‘on’ parametersestimated by fitting the model to experimental data. Figure 2 shows the sensorresponse magnitude and profile (generated from these fitted model parameters)for the five types of sensor responding to a rectangular pulse of (i) ethanolvapour and (ii) toluene vapour in air. The odour pulse introduced to the channelhad a flow velocity of 50 cm sK1, a pulse width of 5 s, vapour concentration of0.05 mol mK3 and was at a temperature of 258C. In general, the sensors’responses are computed from the analyte concentration profile and bilinearsensor model at various locations as the pulse traverses down the channel. (N.B.most commercial e-nose instruments only use the ‘spatial’ data shown in figure 2for odour discrimination; in other words, the different magnitudes of sensorresponse at some moment in time).

Figure 3(a) shows the theoretical concentration profiles at five different points,where sensors (e.g. S11) are located along the channel, when a 5 s ethanol vapourpulse is injected into the artificial olfactory mucosa (test conditions as before,stationary phase thickness is 10 mm). As stated above, the temporal delay iscaused by the analyte partitioning into the stationary phase coating by differingamounts. The actual time at which analytes reach a sensor is in general odour


(a) 20ethanol

PEG

PEVAPCLPSB

PVPH

15

refe

renc

e an

alyt

e pu

lse

10ch

ange

in r

esis

tanc

e (%

)

5

0

(b) 10

8

6

4

2

0

chan

ge in

res

ista

nce

(%)

PEVA

PEG

refe

renc

e an

alyt

e pu

lse

toluene

PVPHPSB

PCL

5 10 15 20 25 30time (s)

35 40 45 50

Figure 2. Simulated responses of PSB-, PEVA-, PEG-, PCL- and PVPH-based resistive sensors.Sensor response (as percentage change in resistance) to 5 s pulse of (a) ethanol vapour and (b)toluene vapour in air against time (s).


dependent and more specifically related to the polarity of the odour molecule.The responses of the sensors located at these points along the channel are shownin figure 3b. The temporal delay from sensor S2 to S39 (measured from when thesignal drops to 50% of its response magnitude) was found to be 38.7 s for ethanoland 340.9 s for toluene vapour.

The results of these simulations show that an artificial mucosa, based on thesechannel dimensions and stationary phase coating, should be able to producespatial–temporal profiles suitable for simple linear odour discrimination.

3. Artificial olfactory mucosa fabrication

Our fabricated artificial olfactory mucosa consisted of 40 discrete polymer/carbon black composite chemoresistive sensors (details are given in table 1), aprinted circuit board (PCB) base, into which the sensors were mounted andconnected, and two different lids (with and without stationary phase coating,fabricated from polyester, RS Components Ltd). The lids’ dimensions were


0.025 S2

S11

S28

S39 S40

0.020

0.015

0.010co

ncen

trat

ion

(mol

em

–3)

0.005

0

(a)

chan

ge in

res

ista

nce

(%)

6

5

4

3

2

1

0

(b)

S11

S2

S28

S39 S40

0 10 20 30time (s)

40 50 60 70

Figure 3. Artificial olfactory mucosa sensor responses to ethanol vapour. (a) Ethanol vapourprofiles of five identical sensors located at five different positions along the GC microchannel. (b)Five different polymer types of sensors responding to ethanol vapour at these locations. The signalfor sensor S2 is much lower than the others because the polymer (PSB) responds much more totoluene than ethanol.


10!200!235 mm with a machined channel (channel of dimensions 2.4 m!0.5 mm!0.5 mm) laid out in a serpentine pattern2. The sensors were fabricatedusing custom 3 00 wafer silicon processing technology within the School ofEngineering, University of Warwick. Each device was 2.5!4.0 mm in size andconsisted of a pair of thin coplanar gold electrodes on a SiO2/Si substrate with anelectrode length of 1.0 mm and an inter-electrode gap of 75 mm.

In our final experimental set-up, 10 different polymers were used, incomparison with the five employed in the computer simulations, in order toenhance even further the system’s ability to discriminate between differentodours. The 10 recipes for these polymers or ‘tunings’ are given in table 3. Thepolymer materials were supplied by Sigma Aldrich (UK) and the carbon black(Black Pearls 2000) material was supplied by Cabot Corporation (USA). Thepolymers were either in powder form or small crystals while the carbon blackcontained nanospheres with diameters varying from 50 to 80 nm. The polymers

2 An exact layout of the channel pattern may be found in Tan (2005).


Table 3. Polymer-composite sensing material recipes for 10 different sensor types.

no. polymerpolymermass (g)

mass of carbonblack (g) solvent (20 ml)

1 poly styrene-co-butadiene (PSB) 0.7 0.175 toluene2 poly ethylene-co-vinyl acetate (PEVA) 1.2 0.300 toluene3 poly ethylene glycol (PEG) 1.2 0.300 ethanol4 poly caprolactone (PCL) 1.2 0.300 toluene5 poly 4-vinyl phenol (PVPH) 1.2 0.300 ethanol6 poly 9-vinylcarbazole (PVC) 0.3 0.300 toluene7 poly vinyl pyrrolidone (PVPD) 0.3 0.300 ethanol8 poly(bisphenol-A-carbonate) (PBA) 0.7 0.175 dichloromethane9 poly(sulphane) (PSF) 0.7 0.175 dichloromethane10 poly(chloro-p-xylylene) (PCX) 1.2 0.300 toluene

polymer coating over opening window

passivation opening for bonding

Figure 4. Photograph of actual odour sensor showing the polymer coating above a pair of electrodes.


were first dissolved in their respective solvent overnight with the aid of amagnetic stirrer (SEA, UK) at an elevated temperature of 508C. Carbon blackwas then added and the mixture sonicated for 10 min using a flask shaker (Griffinand George, UK). The mixture was then deposited onto the sensors using anairbrush (HP-BC Iwata, Japan) controlled with a micro-spraying system (RSprecision liquid dispenser, UK). The deposition area was defined by a mechanicalmask, with a circular hole of 1.0 mm diameter, aligned by a purpose-built X–Ystage. The airbrush was held approximately 10–15 cm away from the mask andseveral passes were made depending on the desired thickness (or resistance). Thisactually resulted in a circular coating of typically 1.5 mm in diameter. Theelectrical resistances of the composite polymer sensors were controlled throughthe deposition process to a value of approximately 5 kU with a typical filmthickness of approximately 20 mm (G25%). Finally, figure 4 shows a photographof an odour sensor with the composite polymer coating clearly visible.

After the sensors’ fabrication, the devices were mounted into slots machinedinto the PCB base and gold wire bonded to copper tracks etched into the top.The sensor placements were identical to those given in table 1, though now thesensors S6–S10, S16–S20, S26–S30 and S36–S40 were replaced with poly(9-vinylcarbazole) (PVC), poly(vinyl phenol) (PVPD), poly(bisphenol-A-carbonate)(PBA), poly(sulphane) (PSF) and poly(chloro-p-xylylene) (PCX), with eachblock of five sensors replaced in this order. The resistance of the sensors wasmeasured with a custom-built electronic circuit in which the sensor was the


1.0

(a)

S39 toluene

S1 toluene

S1 ethanol

S39 ethanol

0.8

0.6no

rmal

ized

res

pons

e m

agni

tude

0.4

0.2

0

(b)

norm

aliz

ed r

espo

nse

mag

nitu

de

1.0

0.8

0.6

0.4

0.2

0

S39 toluene

S1 toluene

S1 ethanol

S39 ethanol

0 50 100 150time (s)

200 250 300

Figure 5. Observed responses of sensor S1 (front of column) and S39 (end of column) for an (a)uncoated and (b) coated channel to pulses of both toluene and ethanol vapour in air. From S39, it isevident that the coated column separates out the vapour components in retention time.


feedback resistor of an operational amplifier circuit in an inverting configuration.The reference input resistor is connected to a K2.5 V reference voltage source inorder to produce a positive output to a 16 bit analogue-to-digital converter (ADC).The retentive layer used to coat the channel, as used in the simulations, wasparylene C. This was deposited using a commercial evaporation technique (PDS2010 Labcoater 2, Specialty Coating Systems, Indianapolis, USA). The machineperforms deposition under a vacuum at room temperature, taking approximately3.5 h to deposit a 10 mm thick film.

4. Experimental results and discussion

The final design of the artificial olfactory mucosa was first tested with bothuncoated and coated lids to demonstrate that the retentive coating selectivelydelayed vapours along the channel. Simple polar and non-polar odours (ethanoland toluene vapour in air, 10 s pulse, test temperature 22G28C, humidity of40G5% relative humidity (RH)) were used to validate the operation of thechannel and to make a direct comparison with the simulations. The results of


0.12

0.10

0.08

0.06

0.04

0.02

0.8

0.6

0.4

0.2

0

0

–0.04

–0.02

chan

ge in

vol

tage

(V

)ch

ange

in v

olta

ge (

V)

0 20 40 60 80 100time (s)

ethanol

toluene

PVPDPCL

PCL

PSB

PEVA

PBAPSF

PEGPVC

PCXPVPDPVPH

PBAPEVA

PSF

PSBPCXPVCPEG

PVPH

(a)

(b)

Figure 6. Observed response of different types of sensors responding to simple analytes, i.e. spatialsignals. Sensor responses to 10 s pulse of (a) ethanol vapour and (b) toluene vapour in air.


these tests are shown in figure 5, where the responses of the sensors at the front(sensor S1) and rear (sensor S39) of the channel are shown for both uncoated(figure 5a) and coated (figure 5b) lids. Comparing the uncoated and coated lidsensor responses for sensor S1, it was observed as expected that both the ethanoland the toluene vapour pulses reached the sensor simultaneously. While forsensor S39, at the end of the channel, for the uncoated lid the ethanol and toluenepulses reached the sensor simultaneously (peaks at tw120 s), though for thecoated channel a 31 s delay was observed between the two analytes. These resultsagree well with the finite-element simulation results given above. The data usedin figure 5 have been normalized to emphasize the temporal time delay, hence thespatial height information is not seen.

In order to evaluate further the artificial mucosa, additional experiments wereperformed to ensure that the system was capable of producing significantdiversity in both spatial and temporal signals. These experiments used both lightsimple odours (ethanol and toluene vapour) and heavier/complex odours(essential oils, i.e. peppermint and vanilla, also milk, cream, banana and50 : 50 mixture by volume of peppermint/vanilla and cream/milk). Figure 6a,bshows the response of the first 10 odour sensors to a 10 s pulse of ethanol andtoluene vapour in air, respectively. The response profiles and magnitude


(c)

td = 45 s

S1 S4 S11 S17 S22 S26 S28 S391.0

0.8

0.6

0.4

0.2

norm

aliz

ed r

espo

nse

mag

nitu

de

0

0 25 50 75 100time (s)

125 150 175 200

(b)

td = 58 s

1.0

0.8

0.6

0.4

0.2

norm

aliz

ed r

espo

nse

mag

nitu

de

0

S1 S3 S11 S12 S26 S29 S37 S39

1.0(a)

td = 98 s

0.8

0.6

0.4

0.2

norm

aliz

ed r

espo

nse

mag

nitu

de

0

S1 S3 S11 S12 S26 S29 S37 S39

Figure 7. Observed temporal signals of sensors responding to simple and complex analytes.(a) Normalized sensor responses to toluene vapour in air. (b) Normalized sensor responses toethanol vapour in air. (c) Normalized sensor response to peppermint essence vapour in air.


information correspond to the spatial information given by our system. Clearly,there is considerable diversity in the response. These experiments wereconducted at a flow rate of 25 ml minK1 and a pulse width of 25 s with fiverepetitions at room temperature (25G28C, 40G5% RH).


Table 4. Spatial and temporal sensor response values for both simple and complex odours. (Thespatial response magnitude RON is defined here as the change in sensor output voltage; and thetemporal response as time taken for the odour pulse to travel along the channel.)

sensornumber toluene ethanol

toluene/ethanol peppermint vanilla

peppermint/vanilla banana milk

spatial (V)

S1 2.814 0.454 1.681 1.114 0.282 0.674 0.140 0.166

S12 1.169 0.218 0.610 0.635 0.102 0.347 0.049 0.052

S19 2.220 0.399 1.098 1.030 1.141 0.547 0.061 0.064

S28 0.360 0.276 0.266 0.556 0.085 0.292 0.108 0.143

S40 0.135 0.103 0.095 0.335 0.133 0.236 0.143 0.158

temporal (s)

S1 12.42 12.66 12.89 6.48 7.41 8.96 8.70 5.43

S12 23.36 16.37 22.45 13.35 12.47 14.92 13.60 13.75

S19 26.22 17.61 24.61 14.95 14.18 15.64 15.94 16.62

S28 33.64 20.53 26.25 16.08 13.41 16.61 25.74 30.41

S40 36.15 21.52 27.19 17.45 13.73 17.82 23.35 24.61


Figure 7 shows the normalized response of three sets of sensor responses topulses of (i) toluene vapour, (ii) ethanol vapour and (iii) peppermint essencevapour, showing temporal information from the system. Eight different sensorresponses along the channel are depicted to illustrate different spatial signals (S1and S39 provide the two extremes). Clearly, the sensors have responded insequence according to their placements. As stated above, the time at which thesensor response magnitude reaches 50% is used for comparing the retention timedelay (temporal delay). The relative retention time between S1 and S39 isapproximately 92 s for toluene vapour, 58 s for ethanol vapour and 45 s forpeppermint essence vapour. Using this temporal information, different analytescan be identified. Comparing these results with those obtained by the finite-element simulation, the general trend is similar. Table 4 gives the spatial and thetemporal delay signals from a set of sensors (S2, S12, S19, S28 and S39) along thecolumn for both the simple and the complex odours. These variables were used ina principal components analysis (PCA) to determine linear separability inmultivariate space. Analyses were performed using the spatial data alone, thetemporal data alone and the combined spatio-temporal data. Figure 8a shows aplot of the first two principal components of the spatial data only. In this case,linear classification is reasonable for all of the analytes except milk and banana.Figure 8b shows a PCA plot of only the temporal data and overall the clusters areless tight. However, when the milk and banana samples do now give separateclusters classes although the variation within each cluster is much larger thanthat from other test analytes. Finally, figure 8c shows the results when both thespatial and the temporal data of the sensors are included in the PCA. It isevident that the general performance is better than the temporal data and closerto that of spatial data. On closer inspection, the clusters for banana and milksamples are now not only separable in multivariate space but also much smaller


0.3

0.2

0.1

0

–0.5

–6– 65 – 60 – 55 – 50 – 45 – 40 – 35 – 30 – 25

–4

–2

0

2

4

6

8

10

12

–6

–4

–2

0

2

4

6

8

10

12

–0.4

–0.3

–0.2

–0.1PC

2PC

2PC

2

toluene

tol + etn

spatial

vanilla

milkpep + van

ethanol banana

peppermint

0–1–2–3– 4–5

– 0.5 – 0.4 – 0.3 – 0.2 – 0.1 0 0.1 0.2 0.3PC1

banana

banana

milk

milk

peppermint

peppermint

ethanol

ethanol

pep + van

vanilla

vanilla

toluene

tol + etn

tol + etn

temporal

pep + van

toluene

spatio-temporal

(a)

(b)

(c)

Figure 8. PCA plots of data recorded from five different odour sensors in the artificial olfactorymucosa. (a) Spatial (response magnitude, DV) data only. (b) Temporal (time at 50% of DV) dataonly. (c) Combined spatio-temporal data for five odour sensors.




in width thus providing a better ratio of cluster separation to cluster width(unlike for the spatial data). In analysis of variance with repeated measures(ANOVAR), this is Snecodor’s F-test and is important in defining the statisticalconfidence level for classifying two clusters as belonging to different classes.Therefore, the results show that the spatio-temporal data, which emulate thenasal chromatography principle exhibited in biological olfactory system, providebetter discrimination than either spatial or temporal signals alone. Further andmore advanced data analysis techniques have been developed and also showimproved performance of the spatio-temporal data and these will be published ina subsequent article.

5. Conclusions

A novel artificial olfactory mucosa has been designed, simulated and fabricatedto mimic the nasal chromatography phenomenon that has been observed in themammalian olfactory system. Our system aims to replicate the biological processwhere spatio-temporal signals are extracted by receptor cells placed below themucus layer within the nasal cavity. Using this concept, a system combining 40polymer-composite sensors of 10 different tunings were placed along a 2.4 m longpolymer-coated channel. The distributed sensor system was tested with bothsimple and complex odours in order to examine its capability to generate bothspatial and temporal signals. The results showed that the ‘e-mucosa’ system canproduce such types of signals for different odours, with temporal delays of up to92 s. Multivariate analysis also verified that spatial and temporal signals could beused individually for odour discrimination, but that the combined spatio-temporal signals gave an even better performance. Current e-nose technologygenerally uses spatial signals and the use of spatial/temporal signals based onan artificial olfactory mucosa has clearly been demonstrated to be superior.Future work will be directed towards the analysis of more challenging odourdiscrimination problems (e.g. odour segmentation) and this will require thedevelopment of more advanced topologies and nonlinear spatio-temporal signalprocessing techniques.

The authors would like to thank the Engineering and Physical Research Council (UK) for financialsupport of this research project (grant no. GR/R37975/01).

References

Fontes, E., Byrne, P., Sundqvist, J., Bosander, P. & Marklund, M. 2001 FEMLAB chemicalengineering user manual (version 2.1). Palo Alto, CA: Comsol.

Gardner, J. W. & Bartlett, P. N. 1999 Electronic noses: principles and applications, pp. 1–245.Oxford, UK: Oxford University Press.

Gardner, J. W., Bartlett, P. N. & Pratt, K. F. E. 1995 Modelling of gas-sensitive conductingpolymer devices. IEE Proc.-Circ. Dev. Syst. 142, 321–333. (doi:10.1049/ip-cds:19952170)

Gardner, J. W., Hines, E. L., Molinier, F., Bartlett, P. N. & Mottram, T. T. 1999 Prediction ofhealth of dairy cattle from breath samples using neural network with parametric model ofdynamic response of array of semiconducting gas sensors. Proc. IEE Sci. Meas. Technol. 146,102–106. (doi:10.1049/ip-smt:19990100)

Goldstein, B. 2001 Sensation and perception, p. 704. 6th edn. Belmont, CA: Wadsworth PublishingCompany.


http://dx.doi.org/doi:10.1049/ip-cds:19952170

http://dx.doi.org/doi:10.1049/ip-smt:19990100


Kent, P. F., Mozell, M. M., Murphy, S. J. & Hornung, D. E. 1996 The interaction of imposed andinherent olfactory mucosal activity patterns and their composite representation in amammalian species using voltage-sensitive dyes. J. Neurosci. 16, 345–353.

Kent, P. F., Mozell, M. M., Youngentob, S. L. & Yurco, P. 2003 Mucosal activity patterns as abasis for olfactory discrimination: comparing behavior and optical recordings. Brain Res. 981,1–11. (doi:10.1016/S0006-8993(03)02512-5)

Lysetskiy, M., Lozowski, A. & Zurada, J. M. 2002 Invariant recognition of spatio-temporalpatterns in the olfactory system model. Neural Process. Lett. 15, 225–234. (doi:10.1023/A:1015773115997)

Malnic, B., Hirono, J., Sato, T. & Buck, L. B. 1999 Combinatorial receptor codes for odors. Cell 96,713–723. (doi:10.1016/S0092-8674(00)80581-4)

Mozell, M. M., Sheehe, P. R., Hornung, D. E., Kent, P. F., Youngentob, S. L. & Murphy, S. J. 1987Imposed and inherent mucosal activity patterns. Their composite representation of olfactorystimuli. J. Gen. Physiol. 90, 625–650. (doi:10.1085/jgp.90.5.625)

Noh, H. S., Hesketh, P. J. & Frye-Mason, G. C. 2002 Parylene C gas chromatographic column forrapid thermal cycling. J. Microelectromech. Syst. 11, 718–725. (doi:10.1109/JMEMS.2002.805052)

Pearce, T. C., Schiffman, S. S., Nagle, H. T. & Gardner, J. W. 2003 Handbook of machine olfaction,pp. 419–577. Dordrecht, The Netherlands: Wiley-VCH.

Quenet, E. B., Horn, D. & Dreyfus, G. 2001 Temporal coding in an olfactory oscillatory model.Neurocomputing 38–40, 831–836. (doi:10.1016/S0925-2312(01)00469-6)

Tan, S. L. 2005 Smart chemical sensing microsystem: towards a nose-on-a-chip. PhD thesis,University of Warwick, Coventry, UK.

Tan, S. L., Covington, J. A. & Gardner, J. W. 2006 Velocity-optimised diffusion for ultra-fastpolymer-based resistive gas sensors. IEE Proc. Sci. Meas. Technol. 153, 64–100. (doi:10.1049/ip-smt:20050077)


http://dx.doi.org/doi:10.1016/S0006-8993(03)02512-5

http://dx.doi.org/doi:10.1023/A:1015773115997

http://dx.doi.org/doi:10.1023/A:1015773115997


http://dx.doi.org/doi:10.1085/jgp.90.5.625

http://dx.doi.org/doi:10.1109/JMEMS.2002.805052

http://dx.doi.org/doi:10.1109/JMEMS.2002.805052




towards an artiﬁcial olfactory mucosa for improved odour classiﬁcation · 2008-10-03 ·...

Documents